Nothing
R/merMod.R
In equatiomatic: Transform Models into 'LaTeX' Equations
Defines functions
create_l1.glmerMod
create_l1.lmerMod
create_l1
create_l1_fixef
convert_matrix
create_onecol_array
create_ranef_structure_merMod
create_vcov_merMod
pull_var
create_greek_matrix
assign_vcov_greek
create_means_merMod
rbind_named
pull_slopes
create_slope_intercept
pull_slope_superscript
pull_intercept
check_interact_vary
rm_crosslev_int_redundancy
pull_cross_var
create_fixef_greek_merMod
assign_higher_levels
assign_l1_greek
pull_superscript
pull_min_level
pull_term_subscript
pred_level_split
vary_higher_subscripts
get_order
#' Provides the order of the levels
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}.
#' @noRd
get_order <- function(rhs_random) {
sort(tapply(rhs_random$original_order, rhs_random$group, min))
}
vary_higher_subscripts <- function(term, rhs_random, lev_omit = NULL) {
splt_random <- split(rhs_random, rhs_random$group)
lev_indexes <- letters[seq_along(splt_random) + 9]
order <- get_order(rhs_random)
splt_random <- splt_random[names(order)]
if (!is.null(lev_omit)) {
lev_indexes <- lev_indexes[-seq_along(grep(lev_omit, names(splt_random)))]
splt_random <- splt_random[-seq_along(grep(lev_omit, names(splt_random)))]
}
out <- rep(NA_character_, length(lev_indexes))
for (i in seq_along(splt_random)) {
if (any(grepl(term, splt_random[[i]]$term))) {
out[i] <- paste0(lev_indexes[i], "[i]")
}
}
paste0(out[!is.na(out)], collapse = ",")
}
#' Create a vector to split by
#' The vector denotes which level the term (for the given row)
#' should be added to in the equation (e.g., l1, l2, l3, etc.)
#' @param rhs_fixed output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "fixed", ]}
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}
#' @return A named vector where the names represent the term and the
#' values represent the level. For example, the output may look similar
#' to \code{c("(Intercept)" = "l1", "wave" = "l1",
#' "grouplow" = "sid", "groupmedium" = "sid")}
#' @noRd
pred_level_split <- function(rhs_fixed, rhs_random) {
order <- get_order(rhs_random)
vapply(rhs_fixed$pred_level, function(x) {
if (length(x) == 0) out <- "l1"
if (length(x) == 1) out <- x
if (length(unique(x)) == 1) out <- unique(x)
if (length(unique(x)) > 1) {
out <- order[x]
out <- names(out[which.max(out)])
}
out
}, FUN.VALUE = character(1))
}
pull_term_subscript <- function(greek_coef, n = 1) {
regex <- paste0("(.+\\{.{", n, "}).+(\\})")
gsub(regex, "\\1\\2", greek_coef)
}
pull_min_level <- function(one_crossdata, one_detected, one_lev, one_var,
order) {
all_vars <- mapply_chr(function(lev, var) {
d <- one_crossdata[[lev]]
out <- d[d$term == var, ]$greek
if (length(out) == 0) {
return("")
}
out
}, one_lev, one_var)
min_lev <- one_lev[which.min(order[one_lev])]
all_vars[names(min_lev)]
}
pull_superscript <- function(slp_preds, detect_cross, cross_data, order) {
order <- c("l1" = 0, order)
levs <- Map(function(predlev, detected) {
predlev[detected]
}, slp_preds$pred_level, detect_cross)
vars <- Map(function(splt, detected) {
splt[detected]
}, slp_preds$split, detect_cross)
ss <- mapply_chr(function(cd, detected, lev, var) {
pull_min_level(cd, detected, lev, var, order)
}, cross_data, detect_cross, levs, vars)
superscripts <- if (length(ss) == 0) {
""
} else {
ifelse(is.na(ss), "", ss)
}
superscripts
}
#' Creates the greek for the level 1 fixed effects
#' @param rhs_fixed output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "fixed", ]}
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}
#' @return A character vector the same length at the number of rows in
#' \code{rhs_fixed} with the greek assigned for those that are at level 1 and
#' \code{NA_character_} otherwise (higher-level predictors).
#' @noRd
assign_l1_greek <- function(rhs_fixed, rhs_random) {
beta_indices <- seq_along(rhs_fixed$l1[rhs_fixed$l1])
if (any(rhs_fixed$term == "(Intercept)" & rhs_fixed$l1)) {
beta_indices <- beta_indices[-length(beta_indices)]
}
l1 <- rep(NA_character_, length(rhs_fixed$term))
l1[rhs_fixed$term == "(Intercept)" & rhs_fixed$l1] <- "\\alpha_{"
l1[rhs_fixed$term != "(Intercept)" & rhs_fixed$l1] <- paste0("\\beta_{", beta_indices)
terms <- rhs_fixed$term[!is.na(l1)]
ss <- vapply(terms, function(x) vary_higher_subscripts(x, rhs_random),
FUN.VALUE = character(1))
l1[!is.na(l1)] <- paste0(l1[!is.na(l1)], ss, "}")
l1
}
#' Takes one level of the fixed effects data and assigns greek for that level
#' @param lev_data One level of the data, e.g. \code{splt[[1]]} where
#' \code{splt} is the \code{rhs_fixed} data split by the \code{predsplit}
#' column.
#' @param lev_data_name A character vector specifying the name of the level,
#' e.g., \code{"sid"}, \code{"school"}, \code{"site"}.
#' @param splt The \code{rhs_fixed} data split by the \code{predsplit}
#' column.
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}
#' @return The full data \code{rhs_fixed} data frame for that level with the
#' \code{greek} column filled in
#' @noRd
assign_higher_levels <- function(lev_data, lev_data_name, splt, rhs_random) {
if (is.null(lev_data)) {
return()
}
split_terms <- split(lev_data, lev_data$crosslevel)
order <- get_order(rhs_random)
# create intercepts
int_preds <- split_terms$`FALSE`
ss <- lapply(int_preds$term, function(x) vary_higher_subscripts(x, rhs_random, lev_data_name))
int_preds$greek <- paste0("\\gamma_{", seq_len(nrow(int_preds)), ss, "}")
# Cross-level interactions
slp_preds <- split_terms$`TRUE`
if (is.null(slp_preds)) {
return(int_preds)
}
detect_cross <- lapply(slp_preds$pred_level, function(x) {
!grepl(lev_data_name, x)
})
cross_data <- Map(function(pred_levs, cross_detected) {
splt[pred_levs[cross_detected]]
},
slp_preds$pred_level,
detect_cross)
superscripts <- pull_superscript(slp_preds, detect_cross, cross_data, order)
superscripts <- pull_term_subscript(superscripts)
coef_number <- unlist(tapply(superscripts, superscripts, seq_along))
subscripts <- vapply(slp_preds$term, function(x) {
vary_higher_subscripts(x, rhs_random, lev_data_name)
}, FUN.VALUE = character(1))
slp_preds$greek <- paste0("\\gamma^{", superscripts, "}_{", coef_number, subscripts, "}")
# return
out <- rbind(int_preds, slp_preds)
out[order(out$original_order), ]
}
#' Create the fixed effects portion of the equation
#' @param model The fitted \code{\link[lme4]{lmer}} model
#' @return The \code{extract_rhs} data frame where \code{effect == fixed} with
#' a new \code{greek} column denoting the greek for the specific
#' coefficient (including indexing for the levels the fixed effects vary
#' at), and a new \code{predsplit} column that allows this data frame
#' to be split by the level at which the variable predicts.
#' @noRd
create_fixef_greek_merMod <- function(model, return_variances) {
rhs <- extract_rhs(model, return_variances)
rhs_fixed <- rhs[rhs$effect == "fixed", ]
rhs_random <- rhs[rhs$effect == "ran_pars", ]
rhs_random <- rhs_random[rhs_random$group != "Residual", ]
order <- get_order(rhs_random)
# fixed effects
rhs_fixed$greek <- assign_l1_greek(rhs_fixed, rhs_random)
rhs_fixed$predsplit <- pred_level_split(rhs_fixed, rhs_random)
splt <- split(rhs_fixed, rhs_fixed$predsplit)[c("l1", names(order))]
splt <- splt[!vapply(splt, is.null, FUN.VALUE = logical(1))]
for (i in seq_along(splt)[-1]) {
splt[[i]] <- assign_higher_levels(splt[[i]], names(splt)[i], splt, rhs_random)
}
Reduce(rbind, splt)
}
pull_cross_var <- function(cross_splt_frame, order) {
if (nrow(cross_splt_frame) == 0) {
return()
}
interaction_terms <- strsplit(cross_splt_frame$term, ":")
order <- c("l1" = 0, order)
orders <- lapply(cross_splt_frame$pred_level, function(x) {
order[x]
})
# put in order from lowest to highest level
interaction_terms <- Map(order_split,
interaction_terms,
cross_splt_frame$pred_level)
term <- mapply_chr(function(interaction, order) {
interaction[which.min(order)]
}, interaction_terms, orders)
unique(term)
}
rm_crosslev_int_redundancy <- function(model, lev_data, term) {
formula_rhs <- labels(terms(formula(model)))
formula_rhs <- formula_rhs[!grepl(":|\\|", formula_rhs)]
terms <- vapply(term, function(x) {
unlist(extract_primary_term(formula_rhs, x))
}, FUN.VALUE = character(1))
subset <- lapply(lev_data$primary, function(x) {
exact <- x %in% paste0(terms, collapse = "|")
detected <- grepl(paste0(terms, collapse = "|"), x)
if (any(exact)) {
!exact
} else {
!detected
}
})
lev_data$primary <- Map(function(x, ss) x[ss], lev_data$primary, subset)
lev_data$subscripts <- Map(function(x, ss) x[ss], lev_data$subscripts, subset)
lev_data
}
# Make cross-level interactions go on the intercept level if the coef
# it predicts doesn't vary at this level (because there's no mean structure
# for it)
check_interact_vary <- function(splt_lev_fixed, splt_lev_random, order) {
cross <- splt_lev_fixed[splt_lev_fixed$crosslevel, ]
lower_var <- pull_cross_var(cross, order)
check_vary <- lower_var %in% splt_lev_random$term
check <- vapply(lower_var, function(x) {
grepl(x, splt_lev_fixed$term)
}, FUN.VALUE = logical(length(splt_lev_fixed$term)))
as.logical(check %*% check_vary)
}
pull_intercept <- function(splt_lev_fixed, splt_lev_random, order,
use_coefs, coef_digits, fix_signs) {
int <- "\\alpha"
if (is.null(splt_lev_fixed)) {
return()
}
check <- check_interact_vary(splt_lev_fixed, splt_lev_random, order)
# Check if lower-level variable in cross-level interactions varies at this level
splt_lev_fixed$crosslevel <- check & splt_lev_fixed$crosslevel
nocross <- splt_lev_fixed[!splt_lev_fixed$crosslevel, ]
# remove any previous superscripts
nocross$greek <- gsub("(.+)\\^.+(_\\{.+$)", "\\1\\2", nocross$greek)
# renumber
nocross$greek <- paste0(gsub("(.+_\\{).+", "\\1", nocross$greek),
seq_along(nocross$greek),
gsub(".+_\\{.{1}(.+)", "\\1", nocross$greek))
# add alpha superscript
nocross$greek <- paste0(nocross$greek, paste0("^{", int, "}"))
if (use_coefs) {
coef_terms <- paste0(round(nocross$estimate, coef_digits),
"_{", nocross$greek, "}",
nocross$terms)
if (length(coef_terms) == 0) {
coef_terms <- 0
}
coef_terms <- paste0(coef_terms, collapse = " + ")
if (fix_signs) {
coef_terms <- fix_coef_signs(coef_terms)
}
out <- data.frame(term = "(Intercept)",
greek = coef_terms,
stringsAsFactors = FALSE)
} else {
coef_terms <- paste0(nocross$greek, nocross$terms)
# add intercept term
coef_terms <- c(paste0("\\gamma_{0}^{", int, "}"), coef_terms)
out <- data.frame(term = "(Intercept)",
greek = paste0(coef_terms, collapse = " + "),
stringsAsFactors = FALSE)
if (nrow(out) == 0) {
return()
}
}
out
}
pull_slope_superscript <- function(greek) {
gsub(".+\\^\\{(.+)\\}_.+", "\\1", greek)
}
create_slope_intercept <- function(term) {
gsub("(.+)(.{1})(\\}$)", "\\10\\3", term)
}
pull_slopes <- function(model, splt_lev_fixed, splt_lev_random, ital_vars,
order, use_coefs, coef_digits, fix_signs,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors) {
if (is.null(splt_lev_fixed)) {
return()
}
check <- check_interact_vary(splt_lev_fixed, splt_lev_random, order)
splt_lev_fixed$crosslevel <- check & splt_lev_fixed$crosslevel
cross <- splt_lev_fixed[splt_lev_fixed$crosslevel, ]
# pull the specific slope that is predicted by the cross-level interaction
cross$slope_predicted <- pull_slope_superscript(cross$greek)
# split the df by the slope that is predicted
cross_splt <- split(cross, cross$slope_predicted)
# pull the name of the variable being predicted
terms_predicted <- lapply(cross_splt, pull_cross_var, order)
# remove that variable from interaction so there's no redundancies
cross_splt <- Map(function(cross, terms) {
rm_crosslev_int_redundancy(model, cross, terms)
}, cross_splt, terms_predicted)
# recreate terms
cross_splt <- lapply(cross_splt, function(x) {
x$terms <- create_term(x, ital_vars, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors)
x$terms <- vapply(x$terms, function(x) {
if (nchar(x) == 0) {
return("")
}
paste0("(", x, ")")
}, character(1))
x
})
if (use_coefs) {
final_slopes <- lapply(cross_splt, function(x) {
slopes <- paste0(round(x$estimate, coef_digits),
"_{",
#gsub("(^.+)\\^\\{.+\\}\\}(.+)",
# "\\1\\2",
cross_splt[[1]]$greek, #),
"}",
x$terms)
paste0(slopes, collapse = " + ")
})
} else {
final_slopes <- lapply(cross_splt, function(x) {
int <- create_slope_intercept(x$greek[1])
slopes <- paste0(x$greek, x$terms)
paste0(c(int, paste0(slopes, collapse = " + ")), collapse = " + ")
})
}
if (fix_signs) {
final_slopes <- lapply(final_slopes, fix_coef_signs)
}
out <- data.frame(term = unlist(terms_predicted),
greek = unlist(final_slopes),
stringsAsFactors = FALSE)
if (nrow(out) == 0) {
return()
}
out
}
rbind_named <- function(l) {
null <- vapply(l, is.null, FUN.VALUE = logical(1))
l <- Map(function(lst, lst_names) {
lst$group <- lst_names
lst
}, l[!null], names(l)[!null])
Reduce(rbind, l)
}
#' Create the mean structure for the VCV matrix
#' @param rhs Output from \code{extract_rhs}
#' @param fixed_greek_mermod Output from \code{create_fixef_greek_merMod}.
#' @param model The fitted \code{\link[lme4]{lmer}} model
#' @param ital_vars Logical, defaults to \code{FALSE}. Should the variable
#' names not be wrapped in the \code{\\operatorname{}} command?
#' @return A data frame with each term that varies at each higher level (term),
#' the level at which it varies (group), the greek for the mean structure
#' (greek), and the greek for the term that is varying (greek_vary). The
#' \code{greek_vary} column defines the left hand side of ~. The greek for the
#' mean structure (greek) will include any group-level predictors. If there are
#' none, it will be replaced by \code{"\\mu_{greek_vary}"} where
#' \code{greek_vary} is actually substituted in.
#' @noRd
create_means_merMod <- function(rhs, fixed_greek_mermod, model, ital_vars,
use_coefs, coef_digits, fix_signs,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors) {
rhs_random <- rhs[rhs$effect == "ran_pars", ]
rhs_random <- rhs_random[rhs_random$group != "Residual" &
!grepl("^cor__", rhs_random$term), ]
rhs_random$term <- gsub("sd__", "", rhs_random$term)
order <- get_order(rhs_random)
fixed_greek_mermod$terms <- create_term(fixed_greek_mermod, ital_vars,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
fixed_greek_mermod$terms <- vapply(fixed_greek_mermod$terms, function(x) {
if (nchar(x) == 0) {
return("")
}
paste0("(", x, ")")
}, character(1))
splt_fixed <- split(fixed_greek_mermod, fixed_greek_mermod$predsplit)
splt_fixed <- splt_fixed[names(order)]
names(splt_fixed) <- names(order)
splt_rand <- split(rhs_random, rhs_random$group)
splt_rand <- splt_rand[names(order)]
names(splt_rand) <- names(order)
ints <- Map(function(fixed, rand) {
pull_intercept(fixed, rand, order, use_coefs, coef_digits, fix_signs)
}, splt_fixed, splt_rand)
ints <- rbind_named(ints)
slopes <- Map(function(fixed, rand) {
pull_slopes(model, fixed, rand, ital_vars, order,
use_coefs, coef_digits, fix_signs,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
}, splt_fixed, splt_rand)
slopes <- rbind_named(slopes)
int_slopes <- rbind(ints, slopes)
if (is.null(int_slopes)) {
out <- rhs_random
out$greek <- NA_character_
} else {
out <- merge(rhs_random, int_slopes,
all.x = TRUE, by = c("group", "term"))
}
out <- out[order(out$original_order), c("group", "term", "greek", "original_order")]
random_vary <- fixed_greek_mermod[fixed_greek_mermod$term %in% unique(out$term), ]
random_vary$greek_vary <- gsub("(.+\\{\\d?\\d?\\d?).+", "\\1}", random_vary$greek)
random_vary$new_order <- random_vary$original_order
out <- merge(out,
random_vary[, c("term", "greek_vary", "new_order")],
by = "term",
all.x = TRUE)
lev_indexes <- setNames(letters[seq_along(order) + 9], names(order))
lev_indexes <- lev_indexes[match(out$group, names(lev_indexes))]
out$greek_vary <- ifelse(
is.na(out$greek_vary),
"",
paste0(gsub("(.+)\\}", "\\1", out$greek_vary), lev_indexes, "}")
)
if (use_coefs) {
out$greek <- ifelse(is.na(out$greek), 0, out$greek)
} else {
out$greek <- ifelse(is.na(out$greek),
paste0("\\mu_{", out$greek_vary, "}"),
out$greek)
}
out[order(out$new_order), ]
}
assign_vcov_greek <- function(rand_lev, means_merMod) {
assign <- lapply(rand_lev$terms, function(x) {
means_merMod$greek_vary[match(x, means_merMod$term)]
})
lapply(assign, function(x) {
if (length(x) == 1) {
return(c(x, x))
} else {
x
}
})
}
create_greek_matrix <- function(v, mat, use_coefs, coef_digits, est) {
if (identical(use_coefs, FALSE)) {
if (length(unique(v)) == 1) {
greek_vcov <- paste0("\\sigma^2_{", v[1], "}")
} else {
greek_vcov <- paste0("\\rho_{",
paste0(v, collapse = ""),
"}",
collapse = "")
}
} else {
greek_vcov <- round(est, coef_digits)
}
mat[v[1], v[2]] <- greek_vcov
mat
}
pull_var <- function(term) {
if (grepl("^cor__", term)) {
ran_part <- gsub("(.+\\.).+", "\\1", term)
term <- gsub(ran_part, "", term, fixed = TRUE)
} else if (grepl("^sd__", term)) {
ran_part <- "sd__"
term <- gsub(paste0("^", ran_part), "", term)
}
term
}
create_vcov_merMod <- function(rhs_random_lev, means_merMod,
use_coefs, coef_digits) {
rand_lev <- rhs_random_lev[, c("group", "term", "estimate")]
rand_lev$terms <- gsub("sd__|cor__", "", rand_lev$term)
rand_lev$terms <- strsplit(rand_lev$terms, "\\.")
# add extra row for reverse on correlation
cors <- rand_lev[grepl("^cor__", rand_lev$term), ]
cors$terms <- lapply(cors$terms, rev)
rand_lev <- rbind(rand_lev, cors)
means <- means_merMod[means_merMod$group == unique(rand_lev$group), ]
rand_lev$vcov_greek <- assign_vcov_greek(rand_lev, means)
rand_lev$terms_var <- vapply(rand_lev$term, pull_var, FUN.VALUE = character(1))
rand_lev <- merge(rand_lev, means[, c("term", "new_order")],
by.x = "terms_var",
by.y = "term",
all.x = TRUE)
rand_lev <- rand_lev[order(rand_lev$new_order), ]
# Create matrix
sd_rows <- grepl("^sd__", rand_lev$term)
mat <- diag(sum(sd_rows))
dimnames(mat) <- list(unique(unlist(rand_lev$vcov_greek[sd_rows])),
unique(unlist(rand_lev$vcov_greek[sd_rows])))
for (i in seq_along(rand_lev$vcov_greek)) {
mat <- create_greek_matrix(rand_lev$vcov_greek[[i]], mat,
use_coefs, coef_digits,
est = rand_lev$estimate[i])
}
mat
}
create_ranef_structure_merMod <- function(model, ital_vars, use_coefs,
coef_digits, fix_signs,
return_variances, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
rhs_random <- rhs[rhs$effect == "ran_pars", ]
order <- get_order(rhs_random[rhs_random$group != "Residual", ])
fixed_greek_mermod <- create_fixef_greek_merMod(model, return_variances)
means_merMod <- create_means_merMod(rhs, fixed_greek_mermod,
model, ital_vars,
use_coefs, coef_digits,
fix_signs, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors)
means_splt <- split(means_merMod, means_merMod$group)[names(order)]
names(means_splt) <- names(order)
rhs_random_splt <- split(rhs_random, rhs_random$group)[names(order)]
vcov_mats <- lapply(rhs_random_splt, function(x) {
create_vcov_merMod(x, means_merMod, use_coefs, coef_digits)
})
# create each final latex piece
lhs <- lapply(means_splt, function(x) create_onecol_array(x$greek_vary))
means <- lapply(means_splt, function(x) create_onecol_array(x$greek))
vcovs <- lapply(vcov_mats, convert_matrix)
distributed <- Map(wrap_normal_dist, means, vcovs)
Map(function(lhs, dist, name, index) {
paste0("\n ", lhs, " \\sim ", dist,
"\n \\text{, for ", escape_tex(name), " ", tolower(index), " = 1,}",
" \\dots ", "\\text{,", toupper(index), "}")
}, lhs, distributed, names(lhs), letters[seq_along(lhs) + 9])
}
#' Create a one column array from a vector
#' This may be able to be incorporated into the `convert_matrix` function
#' @param v character vector to put into a one-column array
#' @keywords internal
#' @noRd
#' @examples \dontrun {
#' equatiomatic:::create_onecol_array(c("\\alpha_{j}", "\\beta_{1j}"))
#'
#' }
#' # Note that the actual function does not return with the linebreak
create_onecol_array <- function(v) {
if (length(v) == 1 | is.null(v)) {
return(v)
}
v <- paste0("&", v)
paste0("\n\\left(\n \\begin{array}{c} \n \\begin{aligned}\n",
paste0(" ", v, collapse = " \\\\\n"),
"\n",
" \\end{aligned}\n",
" \\end{array}\n\\right)\n"
)
}
#' Convert a matrix to a LaTeX array
#' @param mat An R matrix
#' @keywords internal
#' @noRd
#' @examples \dontrun {
#' m <- matrix(c("\\sigma^2_{\\alpha_{j}}",
#' "\\rho \\sigma_{\\beta_{1j}} \\sigma_{\\alpha_{j}}",
#' "\\rho \\sigma_{\\alpha_{j}} \\sigma_{\\beta_{1j}}",
#' "\\sigma^2_{\\beta_{1j}"),
#' ncol = 2)
#' equatiomatic:::convert_matrix(m)
#' }
convert_matrix <- function(mat) {
if (all(dim(mat) == c(1, 1))) {
return(mat)
}
cols <- paste(rep("c", ncol(mat)), collapse = "")
paste0("\n\\left(\n \\begin{array}{", cols, "}\n",
paste(" ", apply(mat, 1, paste, collapse = " & "), collapse = " \\\\ \n"),
"\n \\end{array}\n\\right)\n"
)
}
create_l1_fixef <- function(model, ital_vars, use_coefs, coef_digits,
mean_separate, fix_signs, wrap, terms_per_line,
operator_location, return_variances,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
lhs <- extract_lhs(model, ital_vars, use_coefs, swap_var_names, var_colors)
greek <- create_fixef_greek_merMod(model, return_variances)
terms <- create_term(greek, ital_vars, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors)
terms <- vapply(terms, function(x) {
if (nchar(x) == 0) {
return("")
}
paste0("(", x, ")")
}, character(1))
if (use_coefs) {
sigma <- round(sigma(model), coef_digits)
l1 <- paste0(
round(greek$estimate[greek$predsplit == "l1"], coef_digits),
"_{", greek$greek[greek$predsplit == "l1"], "}",
terms[greek$predsplit == "l1"]
)
} else {
l1 <- paste0(greek$greek[greek$predsplit == "l1"],
terms[greek$predsplit == "l1"])
if(is_exposure_modeled(model)) {
offset <- get_offset(model, ital_vars)
l1 <- c(offset, l1)
}
}
if (wrap) {
if (operator_location == "start") {
line_end <- "\\\\\n&\\quad + "
} else {
line_end <- "\\ + \\\\\n&\\quad "
}
l1 <- split(l1, ceiling(seq_along(l1) / terms_per_line))
if (identical(mean_separate, FALSE)) {
l1 <- lapply(l1, function(x) {
terms_added <- paste0(x, collapse = " + ")
paste0("&", terms_added)
})
l1 <- paste0("\\begin{aligned}\n", paste0(l1, collapse = "\\\\"), "\n\\end{aligned}")
if (fix_signs) {
l1 <- fix_coef_signs(l1)
}
} else {
l1 <- lapply(l1, paste0, collapse = " + ")
l1 <- paste0(l1, collapse = line_end)
if (fix_signs) {
l1 <- fix_coef_signs(l1)
}
}
} else {
l1 <- paste0(l1, collapse = " + ")
if (fix_signs) {
l1 <- fix_coef_signs(l1)
}
}
}
create_l1 <- function(model, ...) {
UseMethod("create_l1", model)
}
#' Create the full fixed-effects portion of an lmerMod
#'
#' @param model A fitted model from \code{\link[lme4]{lmer}}
#' @param ital_vars Logical, defaults to \code{FALSE}. Should the variable
#' names not be wrapped in the \code{\\operatorname{}} command?
#' @param sigma The error term. Defaults to "\\sigma^2".
#' @keywords internal
#' @export
#' @noRd
#' @examples \dontrun{
#' library(lme4)
#' fm1 <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)
#' equatiomatic:::create_fixed_merMod(fm1, FALSE)
#' }
create_l1.lmerMod <- function(model, mean_separate,
ital_vars, wrap, terms_per_line,
use_coefs, coef_digits, fix_signs,
operator_location, sigma = "\\sigma^2",
return_variances, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
lhs <- extract_lhs(model, ital_vars, use_coefs, swap_var_names, var_colors)
l1 <- create_l1_fixef(model, ital_vars, use_coefs, coef_digits,
mean_separate, fix_signs, wrap, terms_per_line,
operator_location, return_variances,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
if (is.null(mean_separate)) {
mean_separate <- sum(rhs$l1) > 3
}
if (mean_separate) {
paste0(lhs, " \\sim ", wrap_normal_dist("\\mu", sigma),
" \\\\\n \\mu &=", l1)
} else {
paste(lhs, "\\sim", wrap_normal_dist(l1, sigma))
}
}
#' @export
#' @noRd
create_l1.glmerMod <- function(model, mean_separate,
ital_vars, wrap, terms_per_line,
use_coefs, coef_digits, fix_signs,
operator_location, sigma = "\\sigma^2",
return_variances, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
lhs <- extract_lhs(model, ital_vars, use_coefs, swap_var_names, var_colors)
l1 <- create_l1_fixef(model, ital_vars, use_coefs, coef_digits,
mean_separate, fix_signs, wrap, terms_per_line,
operator_location, return_variances,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
combo <- paste0(which_family(model), "-", which_link(model))
switch(
combo,
"binomial-logit" = binomial_logit_l1(model, lhs, l1, ital_vars),
"poisson-log" = poisson_log_l1(lhs, l1)
)
}
Try the equatiomatic package in your browser
Any scripts or data that you put into this service are public.
equatiomatic documentation built on Jan. 31, 2022, 1:06 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Defines functions create_l1.glmerMod create_l1.lmerMod create_l1 create_l1_fixef convert_matrix create_onecol_array create_ranef_structure_merMod create_vcov_merMod pull_var create_greek_matrix assign_vcov_greek create_means_merMod rbind_named pull_slopes create_slope_intercept pull_slope_superscript pull_intercept check_interact_vary rm_crosslev_int_redundancy pull_cross_var create_fixef_greek_merMod assign_higher_levels assign_l1_greek pull_superscript pull_min_level pull_term_subscript pred_level_split vary_higher_subscripts get_order
#' Provides the order of the levels
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}.
#' @noRd
get_order <- function(rhs_random) {
sort(tapply(rhs_random$original_order, rhs_random$group, min))
}
vary_higher_subscripts <- function(term, rhs_random, lev_omit = NULL) {
splt_random <- split(rhs_random, rhs_random$group)
lev_indexes <- letters[seq_along(splt_random) + 9]
order <- get_order(rhs_random)
splt_random <- splt_random[names(order)]
if (!is.null(lev_omit)) {
lev_indexes <- lev_indexes[-seq_along(grep(lev_omit, names(splt_random)))]
splt_random <- splt_random[-seq_along(grep(lev_omit, names(splt_random)))]
}
out <- rep(NA_character_, length(lev_indexes))
for (i in seq_along(splt_random)) {
if (any(grepl(term, splt_random[[i]]$term))) {
out[i] <- paste0(lev_indexes[i], "[i]")
}
}
paste0(out[!is.na(out)], collapse = ",")
}
#' Create a vector to split by
#' The vector denotes which level the term (for the given row)
#' should be added to in the equation (e.g., l1, l2, l3, etc.)
#' @param rhs_fixed output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "fixed", ]}
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}
#' @return A named vector where the names represent the term and the
#' values represent the level. For example, the output may look similar
#' to \code{c("(Intercept)" = "l1", "wave" = "l1",
#' "grouplow" = "sid", "groupmedium" = "sid")}
#' @noRd
pred_level_split <- function(rhs_fixed, rhs_random) {
order <- get_order(rhs_random)
vapply(rhs_fixed$pred_level, function(x) {
if (length(x) == 0) out <- "l1"
if (length(x) == 1) out <- x
if (length(unique(x)) == 1) out <- unique(x)
if (length(unique(x)) > 1) {
out <- order[x]
out <- names(out[which.max(out)])
}
out
}, FUN.VALUE = character(1))
}
pull_term_subscript <- function(greek_coef, n = 1) {
regex <- paste0("(.+\\{.{", n, "}).+(\\})")
gsub(regex, "\\1\\2", greek_coef)
}
pull_min_level <- function(one_crossdata, one_detected, one_lev, one_var,
order) {
all_vars <- mapply_chr(function(lev, var) {
d <- one_crossdata[[lev]]
out <- d[d$term == var, ]$greek
if (length(out) == 0) {
return("")
}
out
}, one_lev, one_var)
min_lev <- one_lev[which.min(order[one_lev])]
all_vars[names(min_lev)]
}
pull_superscript <- function(slp_preds, detect_cross, cross_data, order) {
order <- c("l1" = 0, order)
levs <- Map(function(predlev, detected) {
predlev[detected]
}, slp_preds$pred_level, detect_cross)
vars <- Map(function(splt, detected) {
splt[detected]
}, slp_preds$split, detect_cross)
ss <- mapply_chr(function(cd, detected, lev, var) {
pull_min_level(cd, detected, lev, var, order)
}, cross_data, detect_cross, levs, vars)
superscripts <- if (length(ss) == 0) {
""
} else {
ifelse(is.na(ss), "", ss)
}
superscripts
}
#' Creates the greek for the level 1 fixed effects
#' @param rhs_fixed output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "fixed", ]}
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}
#' @return A character vector the same length at the number of rows in
#' \code{rhs_fixed} with the greek assigned for those that are at level 1 and
#' \code{NA_character_} otherwise (higher-level predictors).
#' @noRd
assign_l1_greek <- function(rhs_fixed, rhs_random) {
beta_indices <- seq_along(rhs_fixed$l1[rhs_fixed$l1])
if (any(rhs_fixed$term == "(Intercept)" & rhs_fixed$l1)) {
beta_indices <- beta_indices[-length(beta_indices)]
}
l1 <- rep(NA_character_, length(rhs_fixed$term))
l1[rhs_fixed$term == "(Intercept)" & rhs_fixed$l1] <- "\\alpha_{"
l1[rhs_fixed$term != "(Intercept)" & rhs_fixed$l1] <- paste0("\\beta_{", beta_indices)
terms <- rhs_fixed$term[!is.na(l1)]
ss <- vapply(terms, function(x) vary_higher_subscripts(x, rhs_random),
FUN.VALUE = character(1))
l1[!is.na(l1)] <- paste0(l1[!is.na(l1)], ss, "}")
l1
}
#' Takes one level of the fixed effects data and assigns greek for that level
#' @param lev_data One level of the data, e.g. \code{splt[[1]]} where
#' \code{splt} is the \code{rhs_fixed} data split by the \code{predsplit}
#' column.
#' @param lev_data_name A character vector specifying the name of the level,
#' e.g., \code{"sid"}, \code{"school"}, \code{"site"}.
#' @param splt The \code{rhs_fixed} data split by the \code{predsplit}
#' column.
#' @param rhs_random output from \code{extract_rhs.lmerMod}, subset as
#' \code{rhs[rhs$effect == "ran_pars", ]}
#' @return The full data \code{rhs_fixed} data frame for that level with the
#' \code{greek} column filled in
#' @noRd
assign_higher_levels <- function(lev_data, lev_data_name, splt, rhs_random) {
if (is.null(lev_data)) {
return()
}
split_terms <- split(lev_data, lev_data$crosslevel)
order <- get_order(rhs_random)
# create intercepts
int_preds <- split_terms$`FALSE`
ss <- lapply(int_preds$term, function(x) vary_higher_subscripts(x, rhs_random, lev_data_name))
int_preds$greek <- paste0("\\gamma_{", seq_len(nrow(int_preds)), ss, "}")
# Cross-level interactions
slp_preds <- split_terms$`TRUE`
if (is.null(slp_preds)) {
return(int_preds)
}
detect_cross <- lapply(slp_preds$pred_level, function(x) {
!grepl(lev_data_name, x)
})
cross_data <- Map(function(pred_levs, cross_detected) {
splt[pred_levs[cross_detected]]
},
slp_preds$pred_level,
detect_cross)
superscripts <- pull_superscript(slp_preds, detect_cross, cross_data, order)
superscripts <- pull_term_subscript(superscripts)
coef_number <- unlist(tapply(superscripts, superscripts, seq_along))
subscripts <- vapply(slp_preds$term, function(x) {
vary_higher_subscripts(x, rhs_random, lev_data_name)
}, FUN.VALUE = character(1))
slp_preds$greek <- paste0("\\gamma^{", superscripts, "}_{", coef_number, subscripts, "}")
# return
out <- rbind(int_preds, slp_preds)
out[order(out$original_order), ]
}
#' Create the fixed effects portion of the equation
#' @param model The fitted \code{\link[lme4]{lmer}} model
#' @return The \code{extract_rhs} data frame where \code{effect == fixed} with
#' a new \code{greek} column denoting the greek for the specific
#' coefficient (including indexing for the levels the fixed effects vary
#' at), and a new \code{predsplit} column that allows this data frame
#' to be split by the level at which the variable predicts.
#' @noRd
create_fixef_greek_merMod <- function(model, return_variances) {
rhs <- extract_rhs(model, return_variances)
rhs_fixed <- rhs[rhs$effect == "fixed", ]
rhs_random <- rhs[rhs$effect == "ran_pars", ]
rhs_random <- rhs_random[rhs_random$group != "Residual", ]
order <- get_order(rhs_random)
# fixed effects
rhs_fixed$greek <- assign_l1_greek(rhs_fixed, rhs_random)
rhs_fixed$predsplit <- pred_level_split(rhs_fixed, rhs_random)
splt <- split(rhs_fixed, rhs_fixed$predsplit)[c("l1", names(order))]
splt <- splt[!vapply(splt, is.null, FUN.VALUE = logical(1))]
for (i in seq_along(splt)[-1]) {
splt[[i]] <- assign_higher_levels(splt[[i]], names(splt)[i], splt, rhs_random)
}
Reduce(rbind, splt)
}
pull_cross_var <- function(cross_splt_frame, order) {
if (nrow(cross_splt_frame) == 0) {
return()
}
interaction_terms <- strsplit(cross_splt_frame$term, ":")
order <- c("l1" = 0, order)
orders <- lapply(cross_splt_frame$pred_level, function(x) {
order[x]
})
# put in order from lowest to highest level
interaction_terms <- Map(order_split,
interaction_terms,
cross_splt_frame$pred_level)
term <- mapply_chr(function(interaction, order) {
interaction[which.min(order)]
}, interaction_terms, orders)
unique(term)
}
rm_crosslev_int_redundancy <- function(model, lev_data, term) {
formula_rhs <- labels(terms(formula(model)))
formula_rhs <- formula_rhs[!grepl(":|\\|", formula_rhs)]
terms <- vapply(term, function(x) {
unlist(extract_primary_term(formula_rhs, x))
}, FUN.VALUE = character(1))
subset <- lapply(lev_data$primary, function(x) {
exact <- x %in% paste0(terms, collapse = "|")
detected <- grepl(paste0(terms, collapse = "|"), x)
if (any(exact)) {
!exact
} else {
!detected
}
})
lev_data$primary <- Map(function(x, ss) x[ss], lev_data$primary, subset)
lev_data$subscripts <- Map(function(x, ss) x[ss], lev_data$subscripts, subset)
lev_data
}
# Make cross-level interactions go on the intercept level if the coef
# it predicts doesn't vary at this level (because there's no mean structure
# for it)
check_interact_vary <- function(splt_lev_fixed, splt_lev_random, order) {
cross <- splt_lev_fixed[splt_lev_fixed$crosslevel, ]
lower_var <- pull_cross_var(cross, order)
check_vary <- lower_var %in% splt_lev_random$term
check <- vapply(lower_var, function(x) {
grepl(x, splt_lev_fixed$term)
}, FUN.VALUE = logical(length(splt_lev_fixed$term)))
as.logical(check %*% check_vary)
}
pull_intercept <- function(splt_lev_fixed, splt_lev_random, order,
use_coefs, coef_digits, fix_signs) {
int <- "\\alpha"
if (is.null(splt_lev_fixed)) {
return()
}
check <- check_interact_vary(splt_lev_fixed, splt_lev_random, order)
# Check if lower-level variable in cross-level interactions varies at this level
splt_lev_fixed$crosslevel <- check & splt_lev_fixed$crosslevel
nocross <- splt_lev_fixed[!splt_lev_fixed$crosslevel, ]
# remove any previous superscripts
nocross$greek <- gsub("(.+)\\^.+(_\\{.+$)", "\\1\\2", nocross$greek)
# renumber
nocross$greek <- paste0(gsub("(.+_\\{).+", "\\1", nocross$greek),
seq_along(nocross$greek),
gsub(".+_\\{.{1}(.+)", "\\1", nocross$greek))
# add alpha superscript
nocross$greek <- paste0(nocross$greek, paste0("^{", int, "}"))
if (use_coefs) {
coef_terms <- paste0(round(nocross$estimate, coef_digits),
"_{", nocross$greek, "}",
nocross$terms)
if (length(coef_terms) == 0) {
coef_terms <- 0
}
coef_terms <- paste0(coef_terms, collapse = " + ")
if (fix_signs) {
coef_terms <- fix_coef_signs(coef_terms)
}
out <- data.frame(term = "(Intercept)",
greek = coef_terms,
stringsAsFactors = FALSE)
} else {
coef_terms <- paste0(nocross$greek, nocross$terms)
# add intercept term
coef_terms <- c(paste0("\\gamma_{0}^{", int, "}"), coef_terms)
out <- data.frame(term = "(Intercept)",
greek = paste0(coef_terms, collapse = " + "),
stringsAsFactors = FALSE)
if (nrow(out) == 0) {
return()
}
}
out
}
pull_slope_superscript <- function(greek) {
gsub(".+\\^\\{(.+)\\}_.+", "\\1", greek)
}
create_slope_intercept <- function(term) {
gsub("(.+)(.{1})(\\}$)", "\\10\\3", term)
}
pull_slopes <- function(model, splt_lev_fixed, splt_lev_random, ital_vars,
order, use_coefs, coef_digits, fix_signs,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors) {
if (is.null(splt_lev_fixed)) {
return()
}
check <- check_interact_vary(splt_lev_fixed, splt_lev_random, order)
splt_lev_fixed$crosslevel <- check & splt_lev_fixed$crosslevel
cross <- splt_lev_fixed[splt_lev_fixed$crosslevel, ]
# pull the specific slope that is predicted by the cross-level interaction
cross$slope_predicted <- pull_slope_superscript(cross$greek)
# split the df by the slope that is predicted
cross_splt <- split(cross, cross$slope_predicted)
# pull the name of the variable being predicted
terms_predicted <- lapply(cross_splt, pull_cross_var, order)
# remove that variable from interaction so there's no redundancies
cross_splt <- Map(function(cross, terms) {
rm_crosslev_int_redundancy(model, cross, terms)
}, cross_splt, terms_predicted)
# recreate terms
cross_splt <- lapply(cross_splt, function(x) {
x$terms <- create_term(x, ital_vars, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors)
x$terms <- vapply(x$terms, function(x) {
if (nchar(x) == 0) {
return("")
}
paste0("(", x, ")")
}, character(1))
x
})
if (use_coefs) {
final_slopes <- lapply(cross_splt, function(x) {
slopes <- paste0(round(x$estimate, coef_digits),
"_{",
#gsub("(^.+)\\^\\{.+\\}\\}(.+)",
# "\\1\\2",
cross_splt[[1]]$greek, #),
"}",
x$terms)
paste0(slopes, collapse = " + ")
})
} else {
final_slopes <- lapply(cross_splt, function(x) {
int <- create_slope_intercept(x$greek[1])
slopes <- paste0(x$greek, x$terms)
paste0(c(int, paste0(slopes, collapse = " + ")), collapse = " + ")
})
}
if (fix_signs) {
final_slopes <- lapply(final_slopes, fix_coef_signs)
}
out <- data.frame(term = unlist(terms_predicted),
greek = unlist(final_slopes),
stringsAsFactors = FALSE)
if (nrow(out) == 0) {
return()
}
out
}
rbind_named <- function(l) {
null <- vapply(l, is.null, FUN.VALUE = logical(1))
l <- Map(function(lst, lst_names) {
lst$group <- lst_names
lst
}, l[!null], names(l)[!null])
Reduce(rbind, l)
}
#' Create the mean structure for the VCV matrix
#' @param rhs Output from \code{extract_rhs}
#' @param fixed_greek_mermod Output from \code{create_fixef_greek_merMod}.
#' @param model The fitted \code{\link[lme4]{lmer}} model
#' @param ital_vars Logical, defaults to \code{FALSE}. Should the variable
#' names not be wrapped in the \code{\\operatorname{}} command?
#' @return A data frame with each term that varies at each higher level (term),
#' the level at which it varies (group), the greek for the mean structure
#' (greek), and the greek for the term that is varying (greek_vary). The
#' \code{greek_vary} column defines the left hand side of ~. The greek for the
#' mean structure (greek) will include any group-level predictors. If there are
#' none, it will be replaced by \code{"\\mu_{greek_vary}"} where
#' \code{greek_vary} is actually substituted in.
#' @noRd
create_means_merMod <- function(rhs, fixed_greek_mermod, model, ital_vars,
use_coefs, coef_digits, fix_signs,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors) {
rhs_random <- rhs[rhs$effect == "ran_pars", ]
rhs_random <- rhs_random[rhs_random$group != "Residual" &
!grepl("^cor__", rhs_random$term), ]
rhs_random$term <- gsub("sd__", "", rhs_random$term)
order <- get_order(rhs_random)
fixed_greek_mermod$terms <- create_term(fixed_greek_mermod, ital_vars,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
fixed_greek_mermod$terms <- vapply(fixed_greek_mermod$terms, function(x) {
if (nchar(x) == 0) {
return("")
}
paste0("(", x, ")")
}, character(1))
splt_fixed <- split(fixed_greek_mermod, fixed_greek_mermod$predsplit)
splt_fixed <- splt_fixed[names(order)]
names(splt_fixed) <- names(order)
splt_rand <- split(rhs_random, rhs_random$group)
splt_rand <- splt_rand[names(order)]
names(splt_rand) <- names(order)
ints <- Map(function(fixed, rand) {
pull_intercept(fixed, rand, order, use_coefs, coef_digits, fix_signs)
}, splt_fixed, splt_rand)
ints <- rbind_named(ints)
slopes <- Map(function(fixed, rand) {
pull_slopes(model, fixed, rand, ital_vars, order,
use_coefs, coef_digits, fix_signs,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
}, splt_fixed, splt_rand)
slopes <- rbind_named(slopes)
int_slopes <- rbind(ints, slopes)
if (is.null(int_slopes)) {
out <- rhs_random
out$greek <- NA_character_
} else {
out <- merge(rhs_random, int_slopes,
all.x = TRUE, by = c("group", "term"))
}
out <- out[order(out$original_order), c("group", "term", "greek", "original_order")]
random_vary <- fixed_greek_mermod[fixed_greek_mermod$term %in% unique(out$term), ]
random_vary$greek_vary <- gsub("(.+\\{\\d?\\d?\\d?).+", "\\1}", random_vary$greek)
random_vary$new_order <- random_vary$original_order
out <- merge(out,
random_vary[, c("term", "greek_vary", "new_order")],
by = "term",
all.x = TRUE)
lev_indexes <- setNames(letters[seq_along(order) + 9], names(order))
lev_indexes <- lev_indexes[match(out$group, names(lev_indexes))]
out$greek_vary <- ifelse(
is.na(out$greek_vary),
"",
paste0(gsub("(.+)\\}", "\\1", out$greek_vary), lev_indexes, "}")
)
if (use_coefs) {
out$greek <- ifelse(is.na(out$greek), 0, out$greek)
} else {
out$greek <- ifelse(is.na(out$greek),
paste0("\\mu_{", out$greek_vary, "}"),
out$greek)
}
out[order(out$new_order), ]
}
assign_vcov_greek <- function(rand_lev, means_merMod) {
assign <- lapply(rand_lev$terms, function(x) {
means_merMod$greek_vary[match(x, means_merMod$term)]
})
lapply(assign, function(x) {
if (length(x) == 1) {
return(c(x, x))
} else {
x
}
})
}
create_greek_matrix <- function(v, mat, use_coefs, coef_digits, est) {
if (identical(use_coefs, FALSE)) {
if (length(unique(v)) == 1) {
greek_vcov <- paste0("\\sigma^2_{", v[1], "}")
} else {
greek_vcov <- paste0("\\rho_{",
paste0(v, collapse = ""),
"}",
collapse = "")
}
} else {
greek_vcov <- round(est, coef_digits)
}
mat[v[1], v[2]] <- greek_vcov
mat
}
pull_var <- function(term) {
if (grepl("^cor__", term)) {
ran_part <- gsub("(.+\\.).+", "\\1", term)
term <- gsub(ran_part, "", term, fixed = TRUE)
} else if (grepl("^sd__", term)) {
ran_part <- "sd__"
term <- gsub(paste0("^", ran_part), "", term)
}
term
}
create_vcov_merMod <- function(rhs_random_lev, means_merMod,
use_coefs, coef_digits) {
rand_lev <- rhs_random_lev[, c("group", "term", "estimate")]
rand_lev$terms <- gsub("sd__|cor__", "", rand_lev$term)
rand_lev$terms <- strsplit(rand_lev$terms, "\\.")
# add extra row for reverse on correlation
cors <- rand_lev[grepl("^cor__", rand_lev$term), ]
cors$terms <- lapply(cors$terms, rev)
rand_lev <- rbind(rand_lev, cors)
means <- means_merMod[means_merMod$group == unique(rand_lev$group), ]
rand_lev$vcov_greek <- assign_vcov_greek(rand_lev, means)
rand_lev$terms_var <- vapply(rand_lev$term, pull_var, FUN.VALUE = character(1))
rand_lev <- merge(rand_lev, means[, c("term", "new_order")],
by.x = "terms_var",
by.y = "term",
all.x = TRUE)
rand_lev <- rand_lev[order(rand_lev$new_order), ]
# Create matrix
sd_rows <- grepl("^sd__", rand_lev$term)
mat <- diag(sum(sd_rows))
dimnames(mat) <- list(unique(unlist(rand_lev$vcov_greek[sd_rows])),
unique(unlist(rand_lev$vcov_greek[sd_rows])))
for (i in seq_along(rand_lev$vcov_greek)) {
mat <- create_greek_matrix(rand_lev$vcov_greek[[i]], mat,
use_coefs, coef_digits,
est = rand_lev$estimate[i])
}
mat
}
create_ranef_structure_merMod <- function(model, ital_vars, use_coefs,
coef_digits, fix_signs,
return_variances, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
rhs_random <- rhs[rhs$effect == "ran_pars", ]
order <- get_order(rhs_random[rhs_random$group != "Residual", ])
fixed_greek_mermod <- create_fixef_greek_merMod(model, return_variances)
means_merMod <- create_means_merMod(rhs, fixed_greek_mermod,
model, ital_vars,
use_coefs, coef_digits,
fix_signs, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors)
means_splt <- split(means_merMod, means_merMod$group)[names(order)]
names(means_splt) <- names(order)
rhs_random_splt <- split(rhs_random, rhs_random$group)[names(order)]
vcov_mats <- lapply(rhs_random_splt, function(x) {
create_vcov_merMod(x, means_merMod, use_coefs, coef_digits)
})
# create each final latex piece
lhs <- lapply(means_splt, function(x) create_onecol_array(x$greek_vary))
means <- lapply(means_splt, function(x) create_onecol_array(x$greek))
vcovs <- lapply(vcov_mats, convert_matrix)
distributed <- Map(wrap_normal_dist, means, vcovs)
Map(function(lhs, dist, name, index) {
paste0("\n ", lhs, " \\sim ", dist,
"\n \\text{, for ", escape_tex(name), " ", tolower(index), " = 1,}",
" \\dots ", "\\text{,", toupper(index), "}")
}, lhs, distributed, names(lhs), letters[seq_along(lhs) + 9])
}
#' Create a one column array from a vector
#' This may be able to be incorporated into the `convert_matrix` function
#' @param v character vector to put into a one-column array
#' @keywords internal
#' @noRd
#' @examples \dontrun {
#' equatiomatic:::create_onecol_array(c("\\alpha_{j}", "\\beta_{1j}"))
#'
#' }
#' # Note that the actual function does not return with the linebreak
create_onecol_array <- function(v) {
if (length(v) == 1 | is.null(v)) {
return(v)
}
v <- paste0("&", v)
paste0("\n\\left(\n \\begin{array}{c} \n \\begin{aligned}\n",
paste0(" ", v, collapse = " \\\\\n"),
"\n",
" \\end{aligned}\n",
" \\end{array}\n\\right)\n"
)
}
#' Convert a matrix to a LaTeX array
#' @param mat An R matrix
#' @keywords internal
#' @noRd
#' @examples \dontrun {
#' m <- matrix(c("\\sigma^2_{\\alpha_{j}}",
#' "\\rho \\sigma_{\\beta_{1j}} \\sigma_{\\alpha_{j}}",
#' "\\rho \\sigma_{\\alpha_{j}} \\sigma_{\\beta_{1j}}",
#' "\\sigma^2_{\\beta_{1j}"),
#' ncol = 2)
#' equatiomatic:::convert_matrix(m)
#' }
convert_matrix <- function(mat) {
if (all(dim(mat) == c(1, 1))) {
return(mat)
}
cols <- paste(rep("c", ncol(mat)), collapse = "")
paste0("\n\\left(\n \\begin{array}{", cols, "}\n",
paste(" ", apply(mat, 1, paste, collapse = " & "), collapse = " \\\\ \n"),
"\n \\end{array}\n\\right)\n"
)
}
create_l1_fixef <- function(model, ital_vars, use_coefs, coef_digits,
mean_separate, fix_signs, wrap, terms_per_line,
operator_location, return_variances,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
lhs <- extract_lhs(model, ital_vars, use_coefs, swap_var_names, var_colors)
greek <- create_fixef_greek_merMod(model, return_variances)
terms <- create_term(greek, ital_vars, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors)
terms <- vapply(terms, function(x) {
if (nchar(x) == 0) {
return("")
}
paste0("(", x, ")")
}, character(1))
if (use_coefs) {
sigma <- round(sigma(model), coef_digits)
l1 <- paste0(
round(greek$estimate[greek$predsplit == "l1"], coef_digits),
"_{", greek$greek[greek$predsplit == "l1"], "}",
terms[greek$predsplit == "l1"]
)
} else {
l1 <- paste0(greek$greek[greek$predsplit == "l1"],
terms[greek$predsplit == "l1"])
if(is_exposure_modeled(model)) {
offset <- get_offset(model, ital_vars)
l1 <- c(offset, l1)
}
}
if (wrap) {
if (operator_location == "start") {
line_end <- "\\\\\n&\\quad + "
} else {
line_end <- "\\ + \\\\\n&\\quad "
}
l1 <- split(l1, ceiling(seq_along(l1) / terms_per_line))
if (identical(mean_separate, FALSE)) {
l1 <- lapply(l1, function(x) {
terms_added <- paste0(x, collapse = " + ")
paste0("&", terms_added)
})
l1 <- paste0("\\begin{aligned}\n", paste0(l1, collapse = "\\\\"), "\n\\end{aligned}")
if (fix_signs) {
l1 <- fix_coef_signs(l1)
}
} else {
l1 <- lapply(l1, paste0, collapse = " + ")
l1 <- paste0(l1, collapse = line_end)
if (fix_signs) {
l1 <- fix_coef_signs(l1)
}
}
} else {
l1 <- paste0(l1, collapse = " + ")
if (fix_signs) {
l1 <- fix_coef_signs(l1)
}
}
}
create_l1 <- function(model, ...) {
UseMethod("create_l1", model)
}
#' Create the full fixed-effects portion of an lmerMod
#'
#' @param model A fitted model from \code{\link[lme4]{lmer}}
#' @param ital_vars Logical, defaults to \code{FALSE}. Should the variable
#' names not be wrapped in the \code{\\operatorname{}} command?
#' @param sigma The error term. Defaults to "\\sigma^2".
#' @keywords internal
#' @export
#' @noRd
#' @examples \dontrun{
#' library(lme4)
#' fm1 <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)
#' equatiomatic:::create_fixed_merMod(fm1, FALSE)
#' }
create_l1.lmerMod <- function(model, mean_separate,
ital_vars, wrap, terms_per_line,
use_coefs, coef_digits, fix_signs,
operator_location, sigma = "\\sigma^2",
return_variances, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
lhs <- extract_lhs(model, ital_vars, use_coefs, swap_var_names, var_colors)
l1 <- create_l1_fixef(model, ital_vars, use_coefs, coef_digits,
mean_separate, fix_signs, wrap, terms_per_line,
operator_location, return_variances,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
if (is.null(mean_separate)) {
mean_separate <- sum(rhs$l1) > 3
}
if (mean_separate) {
paste0(lhs, " \\sim ", wrap_normal_dist("\\mu", sigma),
" \\\\\n \\mu &=", l1)
} else {
paste(lhs, "\\sim", wrap_normal_dist(l1, sigma))
}
}
#' @export
#' @noRd
create_l1.glmerMod <- function(model, mean_separate,
ital_vars, wrap, terms_per_line,
use_coefs, coef_digits, fix_signs,
operator_location, sigma = "\\sigma^2",
return_variances, swap_var_names,
swap_subscript_names,
var_colors, var_subscript_colors) {
rhs <- extract_rhs(model, return_variances)
lhs <- extract_lhs(model, ital_vars, use_coefs, swap_var_names, var_colors)
l1 <- create_l1_fixef(model, ital_vars, use_coefs, coef_digits,
mean_separate, fix_signs, wrap, terms_per_line,
operator_location, return_variances,
swap_var_names, swap_subscript_names,
var_colors, var_subscript_colors)
combo <- paste0(which_family(model), "-", which_link(model))
switch(
combo,
"binomial-logit" = binomial_logit_l1(model, lhs, l1, ital_vars),
"poisson-log" = poisson_log_l1(lhs, l1)
)
}
Try the equatiomatic package in your browser
Any scripts or data that you put into this service are public.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.